Fitting Claude into existing workflows

Augment or redesign?

Augmenting means the process keeps its shape and Claude takes over part of one step. The handoffs, the approvals and the systems all stay put. Benefits arrive quickly, adoption is easy because nobody has to learn a new process, and if it does not work you simply stop. Redesigning means the sequence itself changes, because a constraint the process was built around has gone — the bottleneck was that only two people could read the contracts, and now the reading is not the bottleneck. Redesigns produce the larger gains and carry the larger risk, and they need explicit sponsorship because they cut across roles.

A practical rule: augment first, and let the augmentation tell you whether a redesign is warranted. If four weeks of drafting support replies shows that ninety per cent of drafts go out with no edit, you have evidence for changing how the queue is staffed. If half need rewriting, you have learned something much cheaper than a reorganisation would have taught you.

Find the friction, not the glamour

The step worth changing is rarely the interesting one. It is usually the dull, repeated one: rekeying the same information into a second system, reformatting what someone else sent, chasing people for updates, reading the same forty pages every month to find the six lines that changed. Those steps have three things in common — they happen often, they produce text, and nobody defends them.

Connectors are what remove the copy-paste. They let Claude access apps and services directly — Google Drive, Gmail and Google Calendar among them — and they mirror your existing permissions, so Claude cannot reach anything you could not open yourself. On Team and Enterprise plans, owners must enable connectors for the organisation before individuals can authenticate, and admins can restrict what a connector is allowed to do. Two limits matter when you design a workflow around them: Gmail exposes message metadata but not the content of attachments, and connectors are subject to the underlying provider’s rate limits. Managing them is 5.2.

Design the checkpoints deliberately

Every workflow has a point of consequence: the moment an output reaches a customer, a regulator, a payment run or a personnel file. Integration is safe when a competent person stands at that point, and unsafe when the review has been moved somewhere convenient instead. A review that happens after publication is an audit, not a control. A review by someone with no way to check the source is a signature, not a check.

This is also where scope discipline pays. A workflow in which Claude prepares and a person decides is easy to explain to legal, easy to roll back, and easy to audit. A workflow in which Claude decides is a different governance conversation, and it belongs to Domain 6. When to insist on human review is covered in 2.4.

Rolling it out to people who did not ask for it

Integration fails on adoption far more often than on capability. Anthropic’s published story of Advantage Solutions, a retail services company with tens of thousands of frontline staff, describes a deliberate sequence: an AI office reporting into an executive steering committee that met regularly to review initiatives; training that began with the executive leadership team — more than 150 senior leaders, including the chief executive — before cascading downward; a pilot of about 150 people on a three-module curriculum; and more than fifty business-unit champions building solutions inside their own teams. The pilot scaled to thousands of users, and the number of identified use cases roughly doubled.

The reported outcomes are the sort of thing a workflow redesign is for. A compliance check that had taken 20 to 30 minutes on paper became a photograph and an automated check completed in minutes, redirecting more than 70,000 labour hours a year toward training and floor presence. A weekly finance workflow of around ten hours was compressed to under thirty minutes. What is worth copying is not the scale but the order: sponsorship, then leaders, then a measured pilot, then champions, then everyone.

1. Pick one workflow with a willing owner. An enthusiastic team with a boring process beats a sceptical team with an exciting one.
2. Pilot with real work for a fixed period — four weeks is usually enough to see the exceptions.
3. Write down the boundary: what Claude does, what it must never do, who reviews, and what to do when it is wrong.
4. Train the people doing the work, not just the managers who approved it. Show the failure modes, not only the wins.
5. Name a champion in each team who owns the templates, instructions and knowledge for that workflow.
6. Re-measure and publish the result, including what did not improve. The honest number is what buys you the next workflow.

Traps the wrong answers are built from

You should now be able to

• Distinguish augmenting a workflow from redesigning it, and choose which a situation calls for.
• Locate the highest-friction step in an existing process rather than the most visible one.
• Place review at the point of consequence and state what Claude must never do.
• Use connectors, Projects and skills to remove copy-paste, re-explanation and inconsistency.
• Plan a rollout with sponsorship, a measured pilot, champions and an honest re-measure.
• Check plan-level availability of features before designing a workflow around them.